AI Agents for Collections and Recovery Operations

Collections and recovery has an unusual property among back-office processes: doing it efficiently and doing it correctly can pull in opposite directions. Contact the customer more often and you recover more, up to the point where the contact itself becomes the compliance problem. Automate the decisioning and you handle more volume, right up until an automated decision affects someone whose circumstances the system did not understand.

That tension is why collections is a poor candidate for the kind of end-to-end automation applied elsewhere in finance operations — and a very good candidate for a narrower, better-designed use of AI agents. The distinction matters. The collections step inside a commercial order-to-cash process is a chasing exercise between businesses. Consumer and regulated lending collections is a conduct-supervised process, governed in the UK by the FCA’s Consumer Credit Sourcebook and reinforced by the Consumer Duty’s support outcome, with comparable expectations in other European jurisdictions. Those rules, not the available technology, set the boundary.

Where the effort actually goes

Break an arrears case into its stages and the manual load concentrates in places that have nothing to do with judgement:

• Case assembly. Pulling a full picture of the account together from the servicing platform, the payment history, the contact log, prior arrangements, and any correspondence — often across systems that were never designed to be queried as one.
• Income and expenditure review. Reading what the customer submitted — bank statements, budget forms, benefit letters — and getting the figures into a usable form before anyone can assess affordability.
• Arrangement monitoring. Tracking whether an agreed plan is being kept, and noticing early when it is drifting rather than after it has broken.
• Change detection. Spotting that something in the case has moved: a payment pattern that shifts, a disclosure buried in a call note, a returned letter suggesting the customer has moved.
• Correspondence drafting. Producing accurate, appropriately worded letters and messages that reflect the specific state of the account.
• Handover packs. Assembling the evidence bundle when a case moves to a specialist team, an external partner, or a legal process.

Every one of those is document-heavy, repetitive, and determined by its inputs. None of them is the decision about how to treat the customer — and that separation is the whole design.

An agentic workflow that respects the constraint

A workable architecture mirrors those stages and keeps the collections professional positioned where their judgement is required:

A case-preparation agent assembles the account view: balance and arrears position, payment history, prior arrangements and how they performed, contact attempts and outcomes, and any documents on file. Its output is a structured summary with every element linked back to its source record, so the person reviewing it can verify rather than trust.

A document-extraction agent handles submitted income and expenditure evidence — pulling figures from statements and forms into a consistent structure, flagging what is missing or inconsistent, and leaving the affordability assessment itself to the assessor. This is the same extraction, validation, and routing pattern used elsewhere in document-heavy finance work, applied to material that is unusually sensitive.

A monitoring agent watches active arrangements and raises cases where behaviour has changed, rather than waiting for a scheduled review or a missed payment threshold.

A signal-surfacing agent reads the case history for indicators that warrant human attention — a disclosed health or bereavement event, a mention of a change in employment, a pattern suggesting the customer is prioritising this debt over essentials. It surfaces and routes; it does not conclude.

A drafting agent produces correspondence grounded in the actual case state, using approved templates and language, for a person to review and send.

The shape is deliberately conservative: agents prepare, people decide, and nothing reaches the customer without human review.

The lines that should not be crossed

Some capabilities are technically straightforward and still the wrong thing to build.

Autonomous outbound contact. Contact frequency, timing, and channel are precisely where collections conduct is judged. An agent that initiates contact on its own logic will eventually contact someone it should not have, at a moment it should not have, and the firm will own that.

Affordability and forbearance decisions. Whether to accept a reduced payment, suspend interest, or grant a payment holiday is a decision with a duty attached. An agent can assemble the evidence and even model the options; the decision and its recorded rationale belong to a trained person.

Vulnerability determination. Treating this as a classification task misunderstands it. The consequence of a wrong answer falls on someone already in difficulty, and the correct response to a signal is human assessment, not an automated flag that changes handling.

Legal or enforcement escalation. Moving a case toward enforcement is consequential and largely irreversible from the customer’s perspective. Agents can prepare the pack; the decision is a person’s.

The general principle is the one that applies across governed agentic workflows: automate the work that produces evidence, not the decision that acts on it.

Why this workload stays inside the boundary

Collections data is among the most sensitive material a lender holds. It combines financial position with disclosures people make only under pressure — illness, redundancy, relationship breakdown, caring responsibilities. Much of it falls into special-category territory under data protection law, and all of it belongs to customers who have not consented to it being processed anywhere other than by their lender.

Running these workflows on infrastructure the firm controls addresses that directly rather than by contractual assurance. When the models, the retrieval index over case documents, and the audit log all execute inside the organisation’s own environment, arrears case material never leaves the boundary where it is already governed — the same reasoning that drives on-premises AI in financial services generally, with an unusually clear application here.

Governance and the audit trail

Collections is examined — by internal audit, by outsourced-partner oversight, and periodically by the regulator. An AI capability that cannot produce a reviewable record will not survive that examination.

Three things make it reviewable. A recorded division of labour, documented per step, so it is unambiguous what the agent produced and what a person decided. A complete action log covering every retrieval, extraction, and draft, retained long enough to reconstruct how a case was handled. And scoped access, enforced at the data layer, so an agent working a case can read that case and not the wider portfolio.

None of this changes who is accountable for the outcome. It changes how much of an experienced collections professional’s time is spent assembling context instead of exercising the judgement that the rules actually require of them.

How VDF AI supports collections workflows

VDF AI is built for exactly this profile: high-volume, document-heavy, and conduct-sensitive. VDF AI Agents handle case assembly, extraction, monitoring, and drafting under scoped access policy, so each workflow reaches only the accounts it is permitted to see. Retrieval is grounded in the firm’s own servicing systems and case documents through private RAG, so a summary reflects the actual record rather than a plausible reconstruction. Human-approval steps sit on every output that reaches a customer or changes how a case is handled. And because the platform runs inside the firm’s own environment, arrears data and the full audit trail stay within the security boundary the rest of the servicing estate already sits behind.